completely broken room after purge_room/rejoin

[richvdh]

when you purge a room (via the delete room api), we do not clear the in-memory caches. Then, when you rejoin, we have a bunch of the events in the cache, so do not bother to re-persist them.

This leads to very bad brokenness, like state_groups_state referring to events which do not exist.

(A workaround is of course to restart synapse after purging and before rejoining)

[richvdh]

Of course, if we had foreign key constraints, the failure mode would be a lot less insidious :/

[richvdh]

Cache invalidation in this case is a bit tricky, because we don't actually know which events we're purging, and most of the caches in question aren't structured in a way that makes it easy to purge them by room id.

[richvdh]

oh and thanks to #6547 you can't even re-purge your way out of the problem :( now fixed